int8conv_cuda.group_conv arguments mismatch

[tshadley]

Quick note that these calls are mismatched, stride and dilation swapped.

niti/ti_torch.py

Lines 214 to 220 in d097335

	grad_weight = int8conv_cuda.group_conv(input,
	grad_output,
	dilation,
	padding,
	stride,
	in_channels * min_batch)

niti/pytorch/int8conv-extension/int8conv.cpp

Lines 11 to 17 in d097335

	torch::Tensor tensor_core_group_conv(
	torch::Tensor& input,
	torch::Tensor& weight,
	int32_t stride,
	int32_t padding,
	int32_t dilation,
	int32_t groups);

[lgrativol]

@tshadley Hello, have you ever managed to reproduce the results of the paper ? I was trying to do and with the base script, i get 10 -12 % in accuracy.

[tshadley]

@tshadley Hello, have you ever managed to reproduce the results of the paper ? I was trying to do and with the base script, i get 10 -12 % in accuracy.

Yes, on compute 7.5 (Turing). Got 90.46% around epoch 166:

2023-04-11 16:18:49 - INFO - best_prec1: 90.460000 ./results/test
2023-04-11 16:18:49 - INFO - Epoch: 166 Train Prec@1 99.066 Train Loss 0.027 Valid Prec@1 90.460 Valid Loss 0.521 

[lgrativol]

Ahh ok, thanks for the reply !! I'm trying on GPU Ampere, using the last version of the repo. Well, I'm gonna keep trying to debug it. :)